Sunil Venkataram
Back to garden
•
🌰 seedling
Three Waves of AI Opportunity - Unhobbling, Physical Interface, Robotics
type/noteaiinvestingstrategy

Three Waves of AI Opportunity — Unhobbling, Physical Interface, Robotics

Wave 1 — Digital Unhobbling (now)

The term “unhobbling” captures what’s already happening: tasks that were technically possible for years but that nobody had the human cycles to actually do at scale. The most obvious examples:

  • Reading every email, document, and PDF in a large organization and summarizing what matters
  • Monitoring every sensor and log in a distributed system for anomalies worth investigating
  • Reviewing every pull request, contract, or medical chart against a set of policy rules
  • Triaging every incoming request (patient referral, customer support ticket, job application, ad bid) with judgment, not just rules

All of this was technically possible with enough human labor. None of it was economically possible because human labor was too expensive and too slow. AI unhobbles the backlog. The market is enormous, largely untapped, and the substrate is pure bits. No hardware, no actuators, no physical safety concerns.

Why this wave is tractable now: The models can do it, the integration tax is low, the liability surface is manageable, and the buyer has a budget line (the humans currently doing it, or the opportunity cost of not doing it).

Who wins: Vertical software companies that can wrap unhobbling use cases in workflows specific to an industry. See Healthcare Admin Automation - Data Transformation and Vertical Workflows as one example.

Wave 2 — Digital-Physical Interface (soon)

The second wave adds a thin layer of hardware to the stack. Sensors (cameras, microphones, GPS, IoT) feed signals into AI reasoning loops. Actuators (smart home devices, displays, speakers, network commands) close the loop. The hardware is commodity, no novel engineering required, but the integration is still nontrivial.

Examples already emerging:

  • Persistent home agents that monitor cameras, detect events, and take action (lights, messages, alerts). See Claws - Persistent Looping Agents as App Replacement.
  • Industrial monitoring — cameras + vision models watching production lines, detecting defects, halting machines before damage occurs.
  • Voice-first interfaces that actually work. Not the 2015 Alexa/Siri dead-end, but agent loops that can negotiate multi-step tasks verbally.

Why this wave is harder than Wave 1: The failure modes include real physical consequences. A false positive at the unhobbling layer wastes human review time. A false positive at the digital-physical layer might turn off the heat in winter or lock the wrong door. Liability, reliability, and real-world edge cases become first-class concerns.

Why it’s still tractable: The hardware exists and is cheap. The integration work is the hard part, but it’s engineering, not physics.

Timeline estimate: 2-5 years for the category to mature, longer for specific verticals with high safety requirements (healthcare devices, vehicles, industrial safety).

Wave 3 — Full Physical Robotics (far out)

The third wave is robots that manipulate atoms: pick up a cup, assemble a component, walk across uneven terrain, fold a shirt. The gap between this and the previous wave is enormous.

Karpathy’s framing: atoms are roughly 1 million times harder than bits. The reasons compound:

  • Actuator precision — a robot hand needs force control, tactile feedback, and reliability that digital systems never needed to worry about
  • Real-world training data — simulation doesn’t generalize cleanly to the physical world, and real-world data collection is expensive and slow
  • Energy and compute constraints — robots have to carry their own power and compute, unlike cloud-connected digital systems
  • Safety at scale — a physical robot near humans has failure modes that digital systems don’t
  • Hardware unit economics — the robot itself costs thousands to tens of thousands of dollars, not pennies like a cloud API call
  • Regulatory and liability exposure — industrial and consumer robots face the same kind of certification gauntlet as automobiles and medical devices

Why this wave still gets talked about: The long-run prize is enormous. A world with cheap general-purpose manipulation changes construction, elder care, agriculture, and more. But the path from here to there is genuinely hard, and the “we’re almost there” claims of 2024-2026 will likely look premature by 2030.

Timeline estimate: 10+ years for general-purpose humanoid robotics to be useful in unstructured environments. Specialized robotics (warehouse picking, specific manufacturing tasks, autonomous vehicles in constrained environments) will arrive earlier.

Jevons paradox across all three waves

One observation that unifies all three: making a given capability cheaper does not reduce demand for it. It increases demand, because the cheaper capability unlocks use cases that weren’t viable at the previous price point. The textbook example: ATMs were supposed to eliminate bank branches, but banks opened more branches because each branch became cheaper to staff.

The implication for the three waves:

  • Wave 1: Cheaper AI reasoning → more tasks get automated, not fewer jobs in aggregate. Total AI spend grows.
  • Wave 2: Cheaper sensing/actuation → more closed-loop systems get built, not fewer.
  • Wave 3: If/when robotics gets cheap → more physical work gets done, not less.

This is the same argument as TAM of Intelligence is Infinite applied to each wave separately.

What this means for positioning

  • Wave 1 is where 90% of the near-term value is. If your side project or investment thesis doesn’t specifically justify being in Wave 2 or 3, default to Wave 1.
  • Wave 2 is where the asymmetric upside might be. Fewer competitors, harder to execute, potentially larger moats once you’re in.
  • Wave 3 is a research bet. Treat it as you would treat any decade-scale technology bet — small position, long horizon, accept you will be wrong about timing.

Counter-argument to hold in mind

Sahaj Garg argues in Cognitive Automation Accelerates the Robotics Timeline that Wave 3 is NOT 10+ years — it’s closer to 5-10 years, because the primary bottleneck in robotics R&D has always been cognitive labor (designing, analyzing, iterating) and that bottleneck is precisely what’s being automated now. The serial physical-test cycle stays the same length, but the parallel cognitive loop compresses by an order of magnitude. If correct, this collapses the Three Waves timeline and specifically destroys the “physical labor is a permanent refuge for displaced knowledge workers” assumption. Worth tracking as a live counter-argument, not dismissing. See also Jevons Paradox vs Cognitive Displacement - The Unresolved Tension.

Backlinks (3)
Connected Notes